<?xml version="1.0" encoding="utf-8" standalone="yes"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:content="http://purl.org/rss/1.0/modules/content/">
  <channel>
    <title>Credit-Assignment on Njx&#39;Log</title>
    <link>https://njx-njx.github.io/tags/credit-assignment/</link>
    <description>Recent content in Credit-Assignment on Njx&#39;Log</description>
    <image>
      <title>Njx&#39;Log</title>
      <url>https://njx-njx.github.io/blog-cover.jpg</url>
      <link>https://njx-njx.github.io/blog-cover.jpg</link>
    </image>
    <generator>Hugo -- 0.161.1</generator>
    <language>zh</language>
    <lastBuildDate>Mon, 31 Aug 2026 12:55:00 +0800</lastBuildDate>
    <atom:link href="https://njx-njx.github.io/tags/credit-assignment/index.xml" rel="self" type="application/rss+xml" />
    <item>
      <title>7 轮交互失败了该怪谁：奖励归因怎么进 GRPO/GSPO</title>
      <link>https://njx-njx.github.io/posts/turn-level-reward-credit-assignment/</link>
      <pubDate>Mon, 31 Aug 2026 12:55:00 +0800</pubDate>
      <guid>https://njx-njx.github.io/posts/turn-level-reward-credit-assignment/</guid>
      <description>单轮 RL 里 advantage 整条序列共享，从没人为难过；到了多轮 agentic RL，同一条轨迹里第 1 轮搜索做对了、第 2 轮 query 跑偏、最后答错——单一 trajectory reward 该怎么分摊，变成一个真实的设计决策。本文拆解形式化层（action mask、轨迹概率）、归因谱系（ORM/PRM/turn 折扣/三条中间路线）、以及 GSPO-token 为多轮留的那个口子。</description>
    </item>
  </channel>
</rss>
